The day Jensen Huang left Nvidia, @htaneja (CEO, General Catalyst) went on stage at Cerebral Valley Summit and gave the most honest take on AI I’ve heard this year.
9 takeaways 👇
1/ Nvidia is “too big to fail,” like the banks were. More competition wouldn’t be the worst outcome. His advice to scaling companies: no single point of failure. Build so you can run on AMD and other chips too.
2/ China: stop underestimating it. More energy capacity, better manufacturing, and it has commoditized AI through open source. Assume their innovation is as good as ours, and compete on that basis.
3/ The AI backlash is self-inflicted. You can’t announce mass job losses plus a 10% chance of catastrophe and expect people to embrace AI. Sell abundance (cheap healthcare, education, intelligence) and agency (opportunity for everyone, not a handful of companies).
4/ Concentration is the real risk: ~75% of compute goes to 6-7 companies, 43% of VC dollars to 2. “Capitalism is a privilege” in a democracy.
5/ Tech should go back to being non-partisan. GC gives to neither side.
6/ Rethink scale. GC invests $10B+/year. When incumbents are worth trillions, a $1B outcome barely matters. Target $100B, then $1T companies.
7/ Recursive self-improvement is “quasi here.” Models already improve compute efficiency.
8/ The big one: for mainstream use cases, intelligence is becoming a commodity. Token volumes swing between labs with every release. Value moves to relationships and to assets you own.
9/ That’s why he backs buying and transforming incumbents over building AI-native challengers. His example: Tesla, the best operator of a generation, still has ~2% of the car market.
Credit photos : @NewcomerMedia
5
175
Niels Rolland retweeted
Four months after we launched Manus, I was invited to a dinner with President Macron. I ended up seated directly across from Jensen Huang.
He was half an hour late, and Macron asked him to give a toast. Jensen picked up his wine glass and turned it into a fully improvised speech — in front of a table of European politicians and founders. It was electric. The charisma of a top founder, on full display.
I typed the whole thing into my Notes that night. Still one of the sharpest takes I’ve heard on software, AI and Europe.
What a pity. What a loss.
Started the day with this great @htaneja fireside chat at CVAI Summit.
He reminds us that AI needs to benefit all humanity to be accepted and adopted. Capitalism is a privilege!
3
148
Spending the day at @stripe HQ.
Michael (@grinich) from WorkOS stayed a full hour with a few of us founders just answering questions and helping, one by one.
The best founders are the ones who give back!
Thanks @michelleefang for the invite.
1
8
320
I'll be back in San Francisco next week, one year later.
I’ll be at a gathering with some of the best AI founders in the world, and I’d love to make the most of the trip by meeting people working on:
→ AI for logistics and supply chain
→ Open-weight and vertical models
→ Investing in vertical AI
Who should I meet in SF? Tag them
7
205
The ability of large models to create / improve smaller, highly specialized models at dramatically lower costs is something I’m particularly excited about.
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations.
Basically one-shotted a full NN that’s better than training it
5
148
Is @instinct not paying bills or just banned from Meta ? @noahrshinn 👻
1
4
280
While Siri is still figuring out how to search the web
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
3
118
So Vercel is training its own models 👀
Meet Glyph Cluster on AI Gateway.
𝚜𝚝𝚎𝚊𝚕𝚝𝚑/𝚐𝚕𝚢𝚙𝚑-𝚌𝚕𝚞𝚜𝚝𝚎𝚛
• Free while in stealth for a limited time
• Access restricted to paid users
• Use it for coding & agentic knowledge work
• Prompts may be used for model improvement
vercel.com/changelog/glyph-c…
3
127
Impressive
Introducing Griffin, the first model to pass the video Turing test.
48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.
It’s the first Human Interaction Model (HIM).
Readers added context they thought people might want to know
The 48% figure and "video Turing test" claim are from Tavus's own study of 54 one-minute calls, not independently verified or using a standard protocol. Griffin-Lite leads NVIDIA's VideoFDB benchmark on their public leaderboard.
cellcog.ai/blog/tavus-gri…
research.nvidia.com/labs/amri/proj…
tech-ish.com/2026/10/02/tav…
This summer, while OpenAI and Anthropic kept racing for raw capability, the real move happened somewhere else.
Stripe bought OpenRouter for over $7B.
Nvidia signed a $6B deal with Poolside.
And last week, Nvidia reportedly agreed to buy Hugging Face for around $13B.
Three massive checks written into a sector whose entire premise is giving away access to LLMs for free. That isn't philanthropy. It's a bet on where enterprise inference is going to run.
Two numbers to frame the debate.
First: the UK AI Safety Institute puts the best open models 4 to 7 months behind the best closed ones.
Four to seven months. Not three years...
Second: per Ramp, only 6% of companies use an open-weight model today.
The gap between those two numbers is the playing field for the next twelve months.
What I'm not saying: that open weight replaces frontier models. On complex reasoning and long agentic tasks, closed models still hold a real lead. We use them every day and that isn't changing tomorrow.
What I am saying: on a narrow, repetitive, well-scoped business process, a properly instructed open model runs just fine on a private cloud instance, or even on-prem. And at that point, three constraints that kill half of enterprise AI projects disappear at once.
So the right question is no longer "open or closed."
It's: for THIS specific process, which model, on which infrastructure, with what traceability, and who is accountable for it.
Niels Rolland retweeted
Google DeepMind CEO Demis Hassabis steps down to become chair ft.trib.al/R3sWuoE
🙃🙃
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!
🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇
🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!
Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
2
427
Spotted le Chaton Fat on stage at Raise Summit
@arthurmensch @MistralAI
Niels Rolland retweeted
I'm not worried about human extinction by AI. I'm worried about human extinction by humans.
That was my closing thought on the @VivaTech stage last week.
For three centuries, work has been our centre of gravity. It has shaped our economic models, our social identity, our sense of personal purpose. So it is worth asking what happens when AI begins to absorb the work we built our lives around.
OpenAI describes its own mission as building systems that "outperform humans at most economically valuable work." If that ambition is realised, the question stops being technical and becomes civilisational. What are humans for?
There will be a painful period of adjustment, where our economic and political models lag behind the technology.
But I am an optimist. I believe the companies that define the next decade will be the ones using AI not to replace us, but to unlock what is most human in us.
The final frontier is not out there. It is in here. As we work less, we will have to relearn how to play more, how to connect more deeply, how to hold on to our own agency.
The tech frontier is fascinating. But the human frontier is the one that matters. That is what I am investing in.
#RogueVentures
This year VivaTech started a little earlier for us.
This afternoon, we officially launched the AI Agent Hub by @PaatchAI on the Champs-Élysées.
Starting Wednesday morning, we’ll be building AI agents live to solve real business problems.
Real workflows.
Real use cases.
Real results.
Coming to @VivaTech ?
Bring us a process, a bottleneck, or a repetitive task.
We’ll build an AI agent for it.
See you there.
👉 vivatech.paatch.ai
































